fix: wrap validation inference with torch.no_grad() in dreambooth examples

[gambletan]

Summary

Fixes #13124

log_validation() in dreambooth training scripts runs pipeline inference using the same UNet/transformer that is being trained, without torch.no_grad(). With --mixed_precision="fp16", gradients are computed and stored during validation inference. When the training loop resumes, the gradient scaler tries to unscale these FP16 gradients, causing:

ValueError: Attempting to unscale FP16 gradients.

This PR wraps all pipeline inference calls in log_validation() with torch.no_grad() across all 17 dreambooth training scripts, not just the one reported in the issue.

Changes

• train_dreambooth.py
• train_dreambooth_flux.py
• train_dreambooth_sd3.py
• train_dreambooth_lora.py
• train_dreambooth_lora_flux.py
• train_dreambooth_lora_flux2.py
• train_dreambooth_lora_flux2_img2img.py
• train_dreambooth_lora_flux2_klein.py
• train_dreambooth_lora_flux2_klein_img2img.py
• train_dreambooth_lora_flux_kontext.py
• train_dreambooth_lora_hidream.py
• train_dreambooth_lora_lumina2.py
• train_dreambooth_lora_qwen_image.py
• train_dreambooth_lora_sana.py
• train_dreambooth_lora_sd3.py
• train_dreambooth_lora_sdxl.py
• train_dreambooth_lora_z_image.py

Root Cause

Validation runs the pipeline which internally calls the UNet/transformer forward pass. Since torch.no_grad() is not set, autograd tracks operations and creates FP16 gradient tensors. The GradScaler in the training loop then fails because it expects FP32 gradients to unscale.

Test Plan

• Verify train_dreambooth_lora.py with --mixed_precision="fp16" and --validation_prompt no longer crashes with ValueError: Attempting to unscale FP16 gradients
• Verify validation images are still generated correctly (inference output is identical with or without no_grad)